Put visual signals
inside your AI workflow.

Connect a stateless MCP endpoint, then ask your agent to analyze a supplied image or sampled video frames. API keys are shared with REST; the service never exposes another organization’s history.

https://autovision.dev/mcp

Connect Codex

Create a key in Settings and set AUTOVISION_API_KEY in the environment launching your MCP client. Keep the secret out of configuration files and source control. Add this to your Codex MCP configuration:

[mcp_servers.autovision]
url = "https://autovision.dev/mcp"
bearer_token_env_var = "AUTOVISION_API_KEY"

The examples use this server’s address; the hosted service is https://autovision.dev/mcp. Refresh the client and inspect the available tools. See the official Codex MCP guide for current setup.

Analysis and inspection tools

• analyze_image: base64 JPEG/PNG image, optional classify.
• analyze_video_frames: up to 12 timed frames and optional classification.

inspect_scene runs bounded autonomous inspection with explicit AI opt-in. Existing analysis tools also accept enhanced:true with classify:true; false disables inspection and omission inherits organization settings. Results include an advisory review recommendation and actual tool trace. Follow-ups consume additional image quota.

All tools use quota and create private job records. There are no shell commands, device-control tools, URL fetching, or automatic incident mutations.

Individual inspection tools

inspect_region takes supplied frames, frame_index and detection_index, remeasures and crops an existing object region. enhance_low_light remeasures and enhances a dark supplied frame. Neither sends images to Bedrock. assess_scene explicitly delegates a frame to the approved scene, package or activity AI lens. All enforce organization tools/settings, charge initial frames plus a follow-up and retain private job evidence. The same executor powers the central Decider. Standalone tools return tool_complete evidence; the consuming agent owns its cross-call budget and next decision.

Discover and stack lenses

Call list_vision_lenses to discover supported and planned capabilities and approved AI model keys without inference or quota charges. One central Decider chooses enabled tools and accumulates their evidence within the shared budget. Scene, package and activity assessment lenses can use different organization-selected approved models. Lenses do not recursively start agents; the consuming app owns approvals and actions.

Try a prompt

Use AutoVision to analyze the supplied image bytes. Explain measured observations separately from the Bedrock summary. Do not infer identity or criminal intent.

An image attached to a conversation is not automatically passed to MCP. The client needs a supported way to read the image and supply base64 bytes as tool arguments. The service does not accept file paths or URLs. Use small, permissioned images.

AgentCore

The endpoint can be consumed by an AgentCore-hosted agent through an authenticated MCP client. Supply the AutoVision API key through the agent’s secret configuration. This endpoint alone is not an AgentCore Runtime deployment; a separately reviewed CloudFormation runtime deployment and AWS invocation test are still required.